roar-cli 0.3.4

Reproducibility and provenance tracker for ML training pipelines

roar

Run Observation & Artifact Registration

roar tracks data artifacts and execution steps in ML pipelines, enabling reproducibility and lineage queries. roar tracking happens automagically by observing your commands as they run, capturing essential context without requiring you to define a pipeline explicitly.

By identifying files based on their actual content rather than their names, it ensures you can always trace a result back to the exact inputs and code that produced it. This gives you reliable reproducibility and a clear history of your artifacts, all derived naturally from your workflow.

While roar captures your work locally, connecting it to a GLaaS (Global Lineage-as-a-Service) server like glaas.ai allows you to publish your lineage graphs to a shared global registry for easy visualization and collaboration. Now your team can search for any artifact by its hash to see exactly how it was made and generate the precise commands needed to reproduce it on another machine.

Installation

pip install roar-cli
# or with uv
uv pip install roar-cli

Requires Python 3.10+.

For the full prereqs, platform support matrix, tracer-backend setup, macOS SIP notes, and sdist build steps, see the canonical Installation docs page. What's below is a TL;DR.

Platform Support

Platform	Status
Linux x86_64	✅ Full support
Linux aarch64	✅ Full support
macOS	🚧 Experimental (limitations)
Windows	Coming soon

PyPI wheels are published for Linux (x86_64, aarch64) and macOS (x86_64, arm64).

If a matching wheel isn't available, pip install falls through to the source distribution. The sdist ships the Rust tracer source but no pre-built binaries, so it requires a C toolchain (gcc / clang), Rust (rustup), and a few minutes to compile the tracers on first install.

Development Installation

# Clone the repository
git clone https://github.com/treqs/roar.git
cd roar

# One-shot dev install: Python package + Rust tracer binaries
bash scripts/install-dev.sh

scripts/install-dev.sh runs pip install -e ".[dev]" (preferring uv when available) and then builds the Rust tracer binaries (roar-tracer, roar-tracer-preload, roar-tracer-ebpf, roard, roar-proxy) and stages them into roar/bin/. A bare pip install -e . does not build the tracer binaries because they live in separate cargo crates outside the maturin manifest, so roar run would fail with "No tracer binary found" until the script runs. See Building from source below for details and the manual flow.

Quick Start

# Initialize roar in your project
cd my-ml-project
roar init

# Run commands with provenance tracking
roar run python preprocess.py --input data.csv --output features.parquet
roar run python train.py --data features.parquet --output model.pt
roar run python evaluate.py --model model.pt --output metrics.json

Product Telemetry

roar keeps anonymous product telemetry counters by default so maintainers can prioritize reliability and platform support work. Telemetry is local-first: small counters are stored under the XDG cache directory and uploaded opportunistically in a background process. Telemetry never uploads file contents, command arguments, file paths, environment variables, repository names, hostnames, usernames, lineage payloads, or GLaaS auth tokens.

Uploaded payloads are limited to:

• A random install_id, event id, sequence number, and coarse timestamps.
• The installed roar version.
• Coarse platform values: OS family, CPU architecture, Python major/minor, shell name, installer class, and whether the process appears containerized.
• Allowlisted command counters such as run, init, register, and successful or failed roar run outcomes.
• Allowlisted tracer selection counters and coarse feature capability flags.

Inspect the current status and exact next payload preview:

roar telemetry --status
roar telemetry --print

When telemetry.endpoint is unset, roar derives the upload endpoint from the configured GLaaS API URL. For example, glaas.url = "https://api.dev.glaas.ai" uses https://api.dev.glaas.ai/api/v1/telemetry/roar.

Disable telemetry globally or for a single project:

roar telemetry --disable
roar config set telemetry.enabled false

Environment opt-outs always win over saved config:

DO_NOT_TRACK=1 roar run python train.py
ROAR_NO_TELEMETRY=1 roar run python train.py

Telemetry is also suppressed automatically in CI, pytest, and Roar-managed backend worker environments such as Ray and OSMO jobs.

Tracer Backends

roar run relies on a Rust "tracer" binary to observe file I/O. If you see an error like "No tracer binary found", build one of the backends below.

Backends

Backend	Binary	Platforms	Notes
eBPF	roar-tracer-ebpf	Linux	Fastest, but requires permissions and kernel support.
preload	roar-tracer-preload + libroar_tracer_preload	macOS, Linux	Uses DYLD_INSERT_LIBRARIES (macOS) or LD_PRELOAD (Linux). Not compatible with processes that ignore preload env vars (e.g., SIP/hardened runtime on macOS), or fully-static binaries (common with Go).
ptrace	roar-tracer	Linux	Slowest, broadest compatibility on Linux.

Building

cd rust

# eBPF (Linux)
cargo build --release -p roar-tracer-ebpf

# preload (macOS & Linux)
cargo build --release -p roar-tracer-preload

# ptrace (Linux)
cargo build --release -p roar-tracer

Selecting A Backend

By default, roar uses auto mode: prefer eBPF, then preload, then ptrace.

# Show what roar can currently find and whether it looks usable
roar tracer

# Set a default backend (auto|ebpf|preload|ptrace)
roar tracer use preload

# Deep preflight for one backend, with the exact failure cause
roar tracer check ebpf

# One-shot host setup for the eBPF backend (applies CAP_BPF)
roar tracer enable ebpf

macOS Tracing Limitations

On macOS, roar uses the preload backend (DYLD_INSERT_LIBRARIES). macOS System Integrity Protection (SIP) silently blocks library injection for Apple-signed platform binaries — anything under /usr/bin/, /bin/, /sbin/, or /System/. When this happens, roar run will complete successfully but capture no file I/O events.

Affected: /usr/bin/python3, /bin/sh, /usr/bin/ruby, and all other Apple-shipped binaries.

Workaround: Use non-Apple builds of your tools:

# Homebrew
brew install python3
roar run python3 train.py          # Uses /opt/homebrew/bin/python3 — works

# conda / pyenv / nix also work
roar run ~/.pyenv/shims/python train.py

# This will NOT capture file events (SIP blocks it):
roar run /usr/bin/python3 train.py

roar prints a warning when it detects no events were captured from a SIP-protected binary.

Commands

roar init

Initialize roar in the current directory. Creates a .roar/ directory to store the local database and a config.toml with default settings.

roar init           # Initialize, prompt for gitignore
roar init -y        # Initialize and auto-add to gitignore
roar init -n        # Initialize without modifying gitignore

roar run <command>

Run a command with provenance tracking. Roar captures:

• Files read and written
• Git commit and branch
• Execution time and exit code
• Command arguments

roar run python train.py --epochs 10 --lr 0.001
roar run ./scripts/preprocess.sh
roar run torchrun --nproc_per_node=4 train.py

# Re-run a previous DAG step
roar run @2                    # Re-run DAG node 2
roar run @2 --epochs=10        # Re-run with parameter override

roar reproduce <hash>

Reproduce an artifact by tracing its lineage.

# Show the reproduction plan (preview)
roar reproduce abc123de

# Run full reproduction
roar reproduce abc123de --run

# Run without prompts
roar reproduce abc123de --run -y

# Include system packages during setup
roar reproduce abc123de --run --package-sync

# Show all required packages (no truncation)
roar reproduce abc123de --list-requirements

# Reproduce a full lineage/session by its 64-character DAG hash
roar reproduce <lineage-hash> --lineage
roar reproduce <lineage-hash> --lineage --run

Unflagged roar reproduce <hash> continues to default to artifact reproduction. Full reproduction clones the git repository, creates a virtual environment, installs recorded packages, and runs the pipeline steps.

roar build <command>

Run a build step with provenance tracking. Build steps run before pipeline steps during reproduction.

# Compile native extensions
roar build maturin develop --release
roar build make -j4

# Install local packages
roar build pip install -e .

Use for setup that should run before the main pipeline (compiling, installing).

roar auth

Manage SSH-key-based GLaaS registration settings.

roar auth register    # Show SSH public key for registration
roar auth test        # Test connection to GLaaS server
roar auth status      # Show current auth status

To register SSH auth with GLaaS:

1. Run roar auth register to display your public key
2. Sign up at https://glaas.ai where you can paste your public key
3. Run roar auth test to verify

roar login

Authenticate with GLaaS and store your global login state for attributed publishing and project access. By default, roar login starts a browser/device login flow and can then unlock private or project-scoped publication in repositories that use roar scope.

roar login
roar login --force
roar login --token-file ~/.config/roar/auth.json

roar config

View or set configuration options.

roar config list
roar config get <key>
roar config set <key> <value>

Run roar config list to see all available options with descriptions. Common options:

Key	Default	Description
output.track_repo_files	false	Include repo files in provenance
output.quiet	false	Suppress written files report
filters.ignore_system_reads	true	Ignore /sys, /etc, /sbin reads
filters.ignore_package_reads	true	Ignore installed package reads
filters.ignore_torch_cache	true	Ignore torch/triton cache
filters.ignore_tmp_files	true	Ignore /tmp files
glaas.url	https://api.glaas.ai	GLaaS server URL
glaas.web_url	https://glaas.ai	GLaaS web UI URL
registration.public_by_default	false	Default register/put visibility
registration.omit.enabled	true	Enable secret filtering
hash.primary	blake3	Primary hash algorithm
logging.level	warning	Log level (debug, info, warning, error)

roar dag

Display the pipeline DAG for the current session.

roar dag                  # Compact view with colors
roar dag --expanded       # Show all executions including reruns
roar dag --json           # Machine-readable JSON output
roar dag --show-artifacts # Show intermediate artifacts

roar env

Manage persistent environment variables injected into roar run and roar build.

roar env set FOO bar      # Set FOO=bar
roar env get FOO          # Print value of FOO
roar env list             # List all env vars
roar env unset FOO        # Remove FOO

roar log

Display recent job execution history.

roar log                  # Show recent job history

roar scope

Show or change the default publication scope for the current repo. Scopes connect lineage from this repo to your personal space or to a specific org/project so later roar register and roar put commands publish in the right place.

roar scope status
roar scope list
roar scope use private
roar scope use acme/foundation-models
roar scope clear

roar label

Manage local labels for DAGs (sessions), jobs, and artifacts.

# Set labels (patches the current label document)
roar label set dag current owner=alice team=ml
roar label set job @2 phase=train lr=0.001
roar label set artifact ./outputs/model.pt model.name=resnet50 stage=baseline

# Remove labels
roar label unset artifact ./outputs/model.pt stage

# Copy labels from one entity to another
roar label cp job @2 artifact ./outputs/model.pt

# Show current labels
roar label show dag current
roar label show job @2
roar label show artifact ./outputs/model.pt

# Show label history (all versions)
roar label history dag current
roar label history artifact <artifact-hash>

# Sync local user-managed labels to GLaaS
roar label sync
roar label sync job @2
roar label sync artifact ./outputs/model.pt --dry-run

roar diff <ref-a> <ref-b>

Compare the lineage of two artifacts, jobs, steps, or sessions to see what changed in their inputs, parameters, code, environment, or pipeline structure. Use it to identify the likely root cause when two outputs differ.

roar diff ./model.pkl ./model-v2.pkl
roar diff @5 @7 --format dag
roar diff session:current session:<hash>
roar diff @5 @7 --json

Entity targets:

• dag: current or a session hash prefix
• job: step ref (@N or @BN) or job UID
• artifact: file path or artifact hash

Labels are stored locally by default. You can explicitly reconcile current local user-managed labels to GLaaS with roar label sync ..., and labels are also included in lineage registration/publish flows when supported by the configured server.

roar register

Register session, job, step, or artifact lineage with GLaaS.

roar register model.pt              # Register model lineage
roar register --dry-run model.pt    # Preview without registering
roar register -y model.pt           # Skip confirmation prompt
roar register @4                    # Register lineage for DAG step 4
roar register deadbeef              # Register lineage for a local job UID
roar register 7f1e...c9a4           # Register lineage for a tracked artifact hash
roar register 8d7a1f2c...           # Register a whole local session
roar register s3://bucket/run/out   # Register a tracked remote S3 artifact

Supported targets:

• Local artifact path: model.pt, ./outputs/metrics.json
• Tracked artifact hash: primitive or composite
• Local job UID: full UID or unique prefix
• Step reference: @N or @BN
• Local session hash: full hash or unique prefix
• Tracked remote path: s3://...

For bare 8-character hex targets, roar register prefers a matching local job UID before falling back to session-hash-prefix resolution.

To make public publication the default for roar register and roar put:

roar config set registration.public_by_default true

Override per command with --public or --private. Use --anonymous on roar register or roar put to force public anonymous publication even when local GLaaS auth is configured. When public visibility comes from config rather than an explicit flag, roar prints a warning before publishing.

roar put

Upload artifacts to cloud storage and register lineage with GLaaS.

roar put model.pt s3://bucket/models/ -m "Final model"
roar put ./checkpoints/ gs://bucket/run-42/ -m "All checkpoints"
roar put @2 s3://bucket/outputs/ -m "Step 2 outputs"

Options:

• -m, --message — Description of the upload (required)
• --dry-run — Preview without uploading
• --no-tag — Skip git tagging
• --public / --private — Override configured publish visibility
• --anonymous — Force public anonymous registration even when local GLaaS auth is configured

Source formats:

• File path: model.pt, ./data/output.csv
• Directory: ./checkpoints/ (uploads all files recursively)
• Job reference: @2 (uploads outputs from step 2)
• No source: uploads all outputs from the current session

roar get

Download artifacts from cloud storage.

roar get s3://bucket/models/model.pt ./local/
roar get gs://bucket/data/train.csv
roar get https://example.com/weights.pt --hash abc123...
roar get s3://bucket/checkpoints/ ./local/ # Download all files under prefix

Options:

• -m, --message — Annotation for this download
• --hash — Expected BLAKE3 hash (for verification)
• --tag — Create a git tag for this download
• --force — Overwrite existing files
• --dry-run — Preview without downloading

Downloads are registered locally as source nodes in the DAG (outputs only, no inputs). They appear in GLaaS when downstream jobs are registered via roar put or roar register.

roar reset

Start a fresh session. Previous session data is preserved in the database.

roar reset                # Reset with confirmation prompt
roar reset -y             # Reset without confirmation

roar show

Show session, job, or artifact details.

roar show                          # Show active session overview
roar show @1                       # Show details for step 1
roar show @B1                      # Show details for build step 1
roar show a1b2c3d4                 # Show job by UID
roar show ./output/model.pkl       # Show artifact by path

roar status

Show a summary of the active session, including the current DAG hash.

roar status

roar workflow

Generate TReqs workflow YAML from a local session.

roar workflow generate
roar workflow generate .treqs/workflows/train.yaml
roar workflow generate --session 8d7a1f2c --name train

Generated workflows follow the TReqs workflow format: name, optional working_directory, and one YAML key per task in session step order. By default, roar workflow generate uses the active session and writes the workflow under .treqs/workflows/ at the repo root.

roar pop

Remove the most recent job from the active session. Useful for undoing a mistaken roar run or correcting the pipeline before registration.

roar pop              # Pop with confirmation prompt
roar pop -y           # Pop without confirmation (skip prompt)

What it does:

• Removes the last job from the session history
• Deletes output artifacts created by that job (unless they're packages/system files)
• Does not affect the original input files

Concepts

Artifacts

Data files tracked by their content hash (BLAKE3). The same file content always has the same hash, regardless of filename or location.

Jobs

Recorded executions that consume input artifacts and produce output artifacts. Each roar run creates a job record.

Collections

Named groups of artifacts, used for downloaded datasets or upload bundles.

Workflow Example

# Record your pipeline
roar run python preprocess.py
roar run python train.py --epochs 10
roar run python evaluate.py

# Later, reproduce an artifact
roar reproduce <model-hash> --run

Git Integration

Roar automatically captures git metadata:

• Current commit hash
• Branch name
• Repository path

Data Storage

All data is stored locally in .roar/roar.db (SQLite). The database includes:

• Artifact hashes and metadata
• Job records with inputs/outputs
• Hash cache for performance

Add .roar/ to your .gitignore (roar offers to do this during roar init).

GLaaS Server

Roar can register sessions, jobs, steps, and artifacts with a GLaaS (Global Lineage-as-a-Service) server using the roar register command.

Server Setup

# Install with server dependencies
uv pip install -e ".[server]"
# or without uv
pip install -e ".[server]"

# Run the server
glaas-server

# Or with custom host/port
GLAAS_HOST=0.0.0.0 GLAAS_PORT=8080 glaas-server

The server provides:

• REST API for artifact and job registration
• Web UI at / with artifact and job browsers
• Search and filtering by command, GPU, file type, etc.

Client Configuration

# Set the GLaaS server URL
roar config set glaas.url http://localhost:8000

# Show your SSH key (copy to GLaaS web UI)
roar auth register

# Test authentication
roar auth test

[!TIP] Roar activity can be registered without authentication. Unauthenticated registrations are attributed to a public "anonymous" user, but are not guaranteed persistence. For persistent attribution, we recommend setting up roar auth.

Development

Prerequisites

• Python 3.10+
• Rust toolchain (for building the tracer) - install from https://rustup.rs/

Setup

bash scripts/install-dev.sh

The script handles Python install + Rust tracer builds + staging binaries into roar/bin/. See Building from source for what it does and how to run the steps manually.

Building from source

pip install -e . runs maturin develop to build the artifact-hash-py pyo3 extension, but the tracer binaries (roar-tracer*, roard, roar-proxy) are separate cargo packages outside the maturin manifest. The PyPI wheels bundle them under roar/bin/; an editable install does not, and roar run fails until they're built and staged.

The fastest path is scripts/install-dev.sh, which does this:

# 1. Python package (editable, with dev extras)
uv pip install -e ".[dev]"   # or pip install -e ".[dev]"

# 2. Build the per-platform tracer crates
cd rust
# Linux:
cargo build --release \
  -p roar-tracer -p roar-tracer-preload -p roar-tracer-ebpf -p roar-proxy
# macOS:
cargo build --release -p roar-tracer-preload -p roar-proxy

# 3. Stage the built binaries where the editable install looks for them
cd ..
mkdir -p roar/bin
# Linux: install five binaries + the preload .so
install -m 0755 rust/target/release/{roar-tracer,roar-tracer-preload,roar-tracer-ebpf,roard,roar-proxy} roar/bin/
install -m 0755 rust/target/release/libroar_tracer_preload.so roar/bin/
# macOS: install the launcher + the preload .dylib + roar-proxy
# install -m 0755 rust/target/release/{roar-tracer-preload,roar-proxy} roar/bin/
# install -m 0755 rust/target/release/libroar_tracer_preload.dylib roar/bin/

The eBPF tracer (Linux only) needs bpf-linker and a Rust nightly toolchain with rust-src for the BPF probe build:

cargo install bpf-linker
rustup install nightly
rustup component add rust-src --toolchain nightly

scripts/install-dev.sh skips eBPF gracefully when bpf-linker is absent — the other tracers still work.

Verify the install with roar tracer; every backend listed should be ready (or have a clear platform-specific reason it isn't, like perf_event_paranoid=4 (needs <= 1) for eBPF on a hardened kernel).

Running Quality Checks

# Linting
ruff check .

# Format check
ruff format --check

# Type checking
mypy roar

# Run all checks at once
ruff check . && ruff format --check && mypy roar

Running Tests

# Run all tests (excluding those requiring a live GLaaS server)
pytest tests/ -v -m "not glaas and not live_glaas"

# Run with coverage
pytest tests/ -v --cov=roar --cov-report=term-missing -m "not glaas and not live_glaas"

# Run tests in parallel
pytest tests/ -v -n auto -m "not glaas and not live_glaas"

# Run only unit tests (fast)
pytest tests/ -v -m "not integration and not e2e and not glaas and not live_glaas"

License

Apache 2.0